Real-Time Navigation Line Extraction for Safflower Harvesting Robots Using an Improved Detection Transformer

This study addresses the limitations in existing navigation techniques for safflower crop row line extraction, including recognition difficulties and poor adaptability. An algorithm using a deformable attention transformer (DAT)-enhanced real-time detection transformer (RT-DETR) network specifically designed for autonomous navigation of safflower harvesting robots is proposed. DualConv BasicBlock was used to enhance the RT-DETR's backbone network by optimizing the BasicBlock module in Residual Network 18, which significantly improved the computational efficiency and feature integration capabilities. Concurrently, the algorithm integrates the Attention-based Intra-scale Feature Interaction module, which enhances the focus on key image regions and effectively reduces computational resource consumption. The Density-Based Spatial Clustering of Applications with Noise algorithm is also used to effectively cluster safflower corolla centroids. Combining centroid calculations in sparse and dense regions with least squares fitting and an adaptive strategy for selecting regional centroids significantly enhanced the row line fitting accuracy and robustness. The experimental results show that the DAT-enhanced RT-DETR network achieved 95.5% mean average precision in safflower corolla detection tasks. Compared to the initial model, the giga floating point operations and parameters of the enhanced network were reduced by 16.5% and 20.1%, respectively. In safflower crop row line fitting tasks, the average fitting time was 28.09 ms and angular error was 4.13°, which outperforms methods that combine Harris corner detection with least squares. Subsequently, field navigation experiments were conducted, resulting in a lateral deviation of ±0.1m for the robot, thereby satisfying the navigation requirements of the safflower harvesting robot. This study enhances the navigation accuracy of safflower harvesting robots and also provides new theoretical foundations and technical approaches for their autonomous navigation.